Deploying 4 parallel Claude Code agents in isolated git worktrees can divide adapter development work and merge results into a clean baseline

Hypothesis

Deploying 4 parallel Claude Code agents in isolated git worktrees can divide adapter development work and merge results into a clean baseline

Changelog

Date	Summary
2026-04-07	Created during temporal gap audit
2026-03-09	Original experiment

Hypothesis

Deploying 4 parallel Claude Code agents in isolated git worktrees, each assigned to a specific ATS channel, can divide adapter development work and produce mergeable results faster than sequential development. The bet is that long-running agent work is bottlenecked on single-threaded attention, not on compute. If I can give each agent its own code domain (a worktree scoped to one adapter), merge conflicts should stay low and I can hold four lines of work in parallel rather than one.

Method

Four Opus sub-agents were assigned to: Direct Discovery, Direct Application, LinkedIn Worker, and Greenhouse Worker. Each operated in an isolated git worktree branched from the same baseline commit. The orchestrator session monitored progress, checked in on each worktree, and merged changes back when an agent signaled completion on a subtask.

Session discipline mattered. Each sub-agent got a scoped task brief (for example, “raise the Greenhouse submission rate by handling the multi-page resume-upload flow”), the specific files it owned, and explicit boundaries on files it must not touch. The orchestrator was responsible for everything cross-cutting: shared types, the state machine, the database schema, and any file that multiple adapters depend on.

Over 2 days the four agents burned 679 total sessions across 9 worktree directories: 522 in the main orchestrator session plus 157 across the four worktrees.

Results

Confirmed. The parallel work produced adapter improvements that merged cleanly into baseline commit e4f43c5. Merge conflict volume stayed low (small fixups in the orchestrator session, nothing that blocked the merge). The pattern demonstrated that Claude Code sessions could be orchestrated as parallel workers with minimal coordination overhead when given disjoint code domains and clear task boundaries.

Findings

Worktree isolation is the key. The moment two agents share a file, merge pain grows superlinearly. The moment they do not, the orchestrator’s only real job is staging and integrating. This maps directly to classic shared-nothing system design: the units can run independently because they do not talk to each other, only to a central coordinator. Four agents at once was comfortable. Six would have strained the orchestrator’s attention. The sweet spot for human-in-the-loop parallel agent work appears to be three to four concurrent agents with a dedicated orchestrator role.

Next Steps

Formalize the pattern into reusable tooling. The insight (“each agent gets its own worktree and its own code domain”) became the core architecture of the Dakka parallel orchestrator project.

Source

jobs-apply four-agent session logs, 2026-03-09 to 2026-03-10. Merged baseline commit e4f43c5.